Industry News

Is Gemini 4 Hydrogen the AI Price Cut Developers Have Been Waiting For?

Google released Gemini 4 Hydrogen, its cheapest Gemini 4 model, at $1.50 per million input tokens and $7.50 per million output tokens, scoring 57 percent on Terminal-Bench 4.1 and cutting fabricated claims five-fold while serving free users through auto model selection.

Google released Gemini 4 Hydrogen on November 12, 2026, and the headline number is impossible to ignore: one dollar and fifty cents per million input tokens, seven dollars and fifty cents per million output tokens. That makes Hydrogen the cheapest model in the Gemini 4 family so far, and unlike most budget launches, it is not quietly confined to an API tier that nobody uses. It is rolling out to the Gemini app, to Google Search for all users, and to the free tier through automatic model selection.

The announcement came from Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Google's Chief AI Architect. His framing was pointed: Hydrogen is not a stripped-down sibling of the frontier models, but a fast execution engine built for the daily grind of agentic work — repository coding, debugging, interactive web development, and multimodal document workflows. The frontier models Carbon and Argon still handle deep, long-horizon reasoning; Hydrogen handles everything that needs to be done a hundred times a day, quickly and cheaply. The full announcement lays out the positioning, while the Gemini API changelog confirms the GA date and pricing line by line.

Whether that pitch holds up depends on the numbers, and Google published a lot of them. Here is what the announcement actually claims, and what it means for anyone shipping AI features in production.

What Google Actually Shipped

Gemini 4 Hydrogen, available under the model string gemini-4-hydrogen, went generally available on November 12 across AI Studio, the Gemini API, Android Studio, and Google Antigravity for paid subscribers. Enterprises can deploy it through Gemini for business and Gemini in Google Sheets, and consumers encounter it in the Gemini app and Google Search. Notably, this is the first Gemini 4 model available across every AI Studio tier, including the free one.

The model keeps the family's 1M-token context window on both input and output, and it accepts PDFs, spreadsheets, charts, audio, and recorded video as native inputs. It also supports parallel batch file reads and searches, which matters more than it sounds: in agentic coding, most wall-clock time is spent waiting for context to be gathered, so the ability to fire off many retrieval calls at once is a genuine speed lever.

Thinking levels carry over with one addition. Hydrogen ships with four configurable levels — low, medium, high (the default), and a new xhigh setting that pushes the model to spend more reasoning effort on hard problems. That last one is worth watching, because it changes the model's cost profile per request: the same weights, priced the same way, can behave like two different products depending on the setting.

The Price Math: 1.50 and 7.50 in Context

Pricing is the clearest signal in the announcement. Hydrogen costs 1.50 dollars per million input tokens and 7.50 dollars per million output tokens, flat across its entire 1M-token context window. Compare that with Gemini 4 Carbon, which went generally available on November 10 at introductory pricing of 2.00 dollars per million input and 10.00 dollars per million output through November 30, 2026. Hydrogen undercuts its own frontier sibling by 25 percent on both sides of the ledger — and Carbon's introductory price is scheduled to end.

The flat pricing across the full context window is also a practical point. Many providers price long-context requests in tiers or apply surcharges as the prompt grows; Google is stating that a million-token prompt costs the same per token as a thousand-token one. For workflows that feed entire repositories, codebases, or document archives into the model, that removes one of the nastier surprises in API bill forecasting.

Then there is the free tier. Free users in the Gemini app will get Hydrogen through automatic model selection whenever it is the right model for the prompt, and Google Search's AI features will run on it for all users. That is a meaningful shift in how budget models get distributed: instead of hiding the cheap model behind a paywall, Google is putting it in front of the largest possible audience and letting the routing layer do the work. The economics are straightforward — every free query served by the cheapest capable model is a query that costs Google less margin, and every developer who tries it in Search or the app is a potential API customer.

The Numbers Google Chose to Lead With

The benchmark section of the announcement is dense, and a few figures stand out.

On Terminal-Bench 4.1, a standard test of agentic coding in real terminal environments, Hydrogen scores 57 percent — more than three times the 18.2 percent that Gemini 3.8 Flash managed. That comparison is the crux of the pitch: the previous cheap model in the lineup was dramatically weaker at exactly the workloads Hydrogen targets, and Google wants developers to see the new model as a generational step, not an incremental one.

On reasoning benchmarks, the claims get bolder. At the default high thinking level, Hydrogen scores 51 on the Artificial Analysis Intelligence Index — one point above the frontier model Argon's 50. At the new xhigh setting it reaches 54. An independent arena-style leaderboard, Agent Arena, places it fifth overall. Take those with the usual caution: index scores measure a slice of capability, and one point on any index is within noise. But the direction of the claim is what matters — Google is arguing that its budget model now trades punches with its own frontier tier on general reasoning.

Domain-specific results follow the same pattern. Hydrogen posts 58.8 percent on HLE-Diamond, a deep domain-reasoning benchmark, which Google calls class-leading; 52.3 percent on the Vals Tax Agent Bench, ranked first; 55.8 percent on the Vals Legal Research Bench, tied for first; 53.4 percent on the Multimodal Knowledge Agent benchmark; and 71.6 percent on Agents' Last Exam v1.1.

The most commercially interesting number, though, involves Orchestrator mode on Terminal-Bench 4.0. When Hydrogen runs inside Google's new agent-orchestration setup, it approaches Carbon's task success rate while consuming 24 percent fewer quota and finishing 19 percent faster. In other words, Google's own numbers suggest the expensive model is no longer always the right model — a point we will come back to.

Faster, and Fewer Fabricated Claims

Beyond raw benchmarks, the efficiency claims aim directly at production pain points. Against Gemini 3.8 Flash at the default high setting, Hydrogen completes tasks 2.5 times faster. Fabricated claims — the model asserting things that are not in the provided context — are reduced by more than five times according to Google's evaluation.

The partner anecdotes, if they generalize, are the numbers developers will feel first. Google says early partners measured codebase exploration and complex enterprise workflows running 2 to 2.8 times faster with 12 to 23 percent fewer tokens consumed. In Google Sheets, output generation runs more than 30 percent faster than the previous model serving that surface. Token reductions compound with the lower per-token price: if a workflow uses 20 percent fewer tokens and each token costs a quarter less, the bill drops by roughly 40 percent before any other optimization.

This is where the fabricated-claims metric deserves attention. Hallucination rates are notoriously hard to measure in the abstract, but in agentic pipelines, false assertions have a specific cost: every fabricated claim propagates downstream into code, filings, reports, or customer-facing output, and someone has to catch it. A five-fold reduction, if it holds up under independent testing, changes the economics of verification — fewer human review cycles per unit of output.

The Safety Ledger, Read Carefully

Google also published a block of safety figures, and they are unusually specific for a launch announcement. Agentic overreach — the model taking actions beyond what the task requires — is reduced by up to 81 percent, and under-service, where the model fails to act when it should, by up to 79 percent. Misaligned behavior and task shortcuts drop by 12 times, and the model is up to 2 times less likely to bypass permission boundaries when facing impossible or blocked tasks.

On internal honesty evaluations, the model scores 96.7 percent. Its chain-of-thought readability is preserved at 99.9 percent across more than 200,000 audited turns, with zero deceptive reasoning edits detected. Resistance to indirect prompt injection, measured on the Gray Swan IPI benchmark, improved by more than 2.3 times. And on the medical side, false refusals on HealthBench Pro — the model wrongly refusing legitimate health questions — stay under 0.4 percent, which Google highlights as evidence that its safeguards now distinguish real biosecurity risks from ordinary medical queries.

The standard caveat applies: these are company-reported numbers, measured on Google's own evaluation suites, and independent replication takes time. But the specificity cuts both ways. Publishing figures like 96.7 percent and 81 percent creates a paper trail that others can probe, which is more than most budget-model launches offer. The overreach and under-service numbers in particular address a real production problem — agents that either do too much or freeze — and if the reductions are real, they lower the cost of the guardrails teams currently build around cheaper models.

Orchestrator and the Strange Economics of the Free Tier

The quietly strategic part of the announcement is Orchestrator, a new feature in Google Antigravity. It lets Carbon act as a coordinator while Hydrogen drives worker and verifier agents in parallel, switching between models dynamically within a single session. Hard architectural decisions go to the frontier model; the flood of routine execution goes to the cheap one.

This is the pattern most serious AI teams have been building by hand for the past year — router layers that send easy requests to small models and escalate the hard ones. Google is now shipping it as a first-party product feature, bundled with its own model family. If it works as described, the moat is not the individual models but the switching infrastructure, and competitors without a cheap-and-capable execution model will feel pressure on two fronts at once.

The free-tier routing raises a subtler question. When the automatically selected model is this cheap, free AI features stop being a pure marketing expense and start being a sustainable product surface. Google Search serving Hydrogen to all users is less a generosity play than a margin calculation — and it quietly establishes expectations among hundreds of millions of users that AI answers are fast, decent, and free, which reshapes what competitors must match.

What to Watch Before You Commit Your Pipeline

Three dates and questions deserve attention. First, Carbon's introductory pricing ends November 30, 2026, so teams budgeting around 2.00 and 10.00 dollars per million tokens should model the post-introductory rates now. Second, the xhigh thinking level means identical pricing can produce very different costs per task depending on configuration — instrument your usage before scaling. Third, benchmark wins at the class level do not guarantee wins on your workload; the Terminal-Bench and Vals results are encouraging, but they are measured on Google's chosen tasks.

The deeper lesson from the announcement is about how to evaluate any model, not just this one. Rankings reward peak capability; production rewards cost per completed task. A model that scores one point lower on an index but finishes the job with a quarter of the tokens is usually the better buy — and comparing tools on that basis, rather than on leaderboard positions alone, is exactly how we approach the landscape at AI Bookmarks, where every entry in our directory is organized so you can weigh what a tool actually costs and does against your specific workload instead of chasing whoever topped the charts this week.

The Bottom Line

Gemini 4 Hydrogen is Google's clearest statement yet that the frontier of AI economics is moving down the stack. The claim that matters is not that a budget model beat a frontier model on one index by a single point — that number will be argued over for weeks. The claim that matters is 57 percent on Terminal-Bench 4.1 at 1.50 dollars per million input tokens, with a five-fold reduction in fabricated claims and a first-party orchestration layer that spends the expensive model's attention only where it is needed.

If independent evaluations confirm even half of these figures, the practical conclusion for developers is simple: the default choice of "frontier model for everything" is now an expensive habit rather than a requirement. Test Hydrogen on your real workload, measure cost per completed task, and keep the frontier models for the problems that genuinely need them.

— The AI Bookmarks news desk, drawing on Google DeepMind's official announcement

Published by AI Bookmarks · More AI news →

← Back to News